The hiring failure everybody talks about is the bad hire. It is visible, it is painful, and you learn from it.
The expensive one is quieter: a good candidate applied, and nobody got to them in time. They took another offer while their CV sat in a stack. You never found out they were good, because nobody read far enough to tell.
That loss leaves no trace. There is no rejected candidate to review, no bad outcome to investigate, no row in any system saying this one was worth calling and the call came four days late. It looks exactly like a week where nothing much came in.
Why speed decides more than judgement
In a competitive market the first decent conversation often wins. Not the best process — the first one. Someone who is genuinely good is being spoken to by other people, and their decision gets made inside a window that opens when they apply and closes when somebody else makes them an offer.
So the question is not how well do you assess candidates? It is how long does a CV sit before a human looks at it?
For most agencies and most employers the honest answer is a few days, and it is not negligence. It is arithmetic. A role gets a pile of applications. Most are not a fit. Reading them is tedious, so it happens in batches, when someone has a clear hour — which is to say, after the window has started closing.
The pile is the problem, and the pile is the thing software can genuinely help with. But only one specific kind of help, and the distinction matters enough that most of this post is about it.
The version I will not build
Automatic rejection. A system that scores applicants and declines the low ones without a person seeing them.
I want to be direct about why, because it is the version people ask for and it is cheaper to build.
The bias problem is real and it is not solved by removing names. A model trained or prompted on your past hiring decisions learns your past hiring decisions, including the ones you would not defend. Proxies survive anonymisation easily — a postcode, a school, a gap in employment, the phrasing someone uses when English is their second language. The system does not need a protected characteristic to discriminate on it; it needs something correlated with one, and almost everything is.
Automatic rejection makes it unauditable. If a person declines a candidate, there is a decision and somebody who made it. If a threshold declines a candidate at three in the morning, there is a number, and when you are asked to explain the pattern across a year of hiring you will not be able to. Several jurisdictions are actively legislating on exactly this — automated employment decision tools, bias audits, candidate notification — and the direction of travel is clear even where the law has not landed yet.
It fails silently and in the expensive direction. A system that wrongly rejects good candidates produces no complaints and no errors. Its output looks identical to a system working perfectly: a smaller pile. You would never know.
So: no automatic rejection. Not as a setting, not as a threshold, not behind a confirmation dialog.
What the useful version does
It changes the order of the pile and shows its reasoning. That is all, and it is enough.
It ranks against requirements you wrote down. Not "good fit" — the specific, stated requirements for that role: the licence, the years in the trade, the shift availability, the location. Written before the applications arrive, which is also just better hiring practice.
It shows why, per candidate, in a line you could read aloud. Matched: CDL class A, three years tanker. Not found: hazmat endorsement. A reason you can check against the CV in front of you. If the reason is wrong you can see that it is wrong, which is the entire property that makes the tool safe.
Everyone stays in the pile. Nothing is removed, nothing is hidden, nothing is declined. Ranking is a reading order, and a reading order you can re-sort.
It flags what it could not determine rather than scoring it zero. A CV that does not mention the licence is unknown, not unqualified. Collapsing those two into one number is how a good candidate with a badly written CV disappears, and it is the same mistake as a system that guesses a plausible value instead of admitting it cannot read one.
A person decides, every time, on every candidate. The tool's job finishes when it has told you where to start reading.
The gain is not fewer decisions. It is that the first hour of attention lands on the candidates most likely to deserve it, on the day they applied rather than next Tuesday. That alone moves the thing that actually decides outcomes.
How would you know if it were working?
Measure the lag, not the model.
Time from application to first human read. This is the number. If it does not drop, nothing else you measure matters.
Time to first contact for candidates you went on to place. The ones that worked are the ones to study.
How often the ranking was wrong — specifically, how often somebody you placed started low in the order. If that never happens the ranking is probably just re-stating the CV's formatting quality. If it happens constantly the requirements were written badly, which is useful to learn.
Who you are hearing from, over time. If the mix of people reaching interview narrows after you introduce ranking, stop and look at why. This is the check that catches the problem the other three cannot.
That last one needs saying plainly: a ranking tool nobody audits for drift is a bias tool with a delay on it.
Where to start
Do not start with software. Start by writing the requirements for one role properly — specific, checkable, decided in advance. Half the value of ranking comes from having done that, and it costs nothing.
Then, if the pile is genuinely the bottleneck, an LLM Integration Sprint — one production feature, end to end, two to four weeks — starts from $7,500. The first conversation is about whether your bottleneck is really the pile, because often it is the vacancy brief, or the response time of the hiring manager, and no tool fixes those.
Before any of that, my website check is free and needs no signup — if candidates apply through your own site, it is worth knowing whether that form actually works on a phone. A broken application form generates no complaints, just an empty pile that looks like a quiet month.
Then the 30-minute Operations Audit, also free: tell me how long a CV currently sits, and I will tell you honestly whether this is worth automating. Get in touch.
